Update libs/rmvpe.py
Browse files- libs/rmvpe.py +670 -670
libs/rmvpe.py
CHANGED
@@ -1,670 +1,670 @@
|
|
1 |
-
from io import BytesIO
|
2 |
-
import os
|
3 |
-
from typing import List, Optional, Tuple
|
4 |
-
import numpy as np
|
5 |
-
import torch
|
6 |
-
|
7 |
-
from
|
8 |
-
|
9 |
-
try:
|
10 |
-
# Fix "Torch not compiled with CUDA enabled"
|
11 |
-
import intel_extension_for_pytorch as ipex # pylint: disable=import-error, unused-import
|
12 |
-
|
13 |
-
if torch.xpu.is_available():
|
14 |
-
from infer.modules.ipex import ipex_init
|
15 |
-
|
16 |
-
ipex_init()
|
17 |
-
except Exception: # pylint: disable=broad-exception-caught
|
18 |
-
pass
|
19 |
-
import torch.nn as nn
|
20 |
-
import torch.nn.functional as F
|
21 |
-
from librosa.util import normalize, pad_center, tiny
|
22 |
-
from scipy.signal import get_window
|
23 |
-
|
24 |
-
import logging
|
25 |
-
|
26 |
-
logger = logging.getLogger(__name__)
|
27 |
-
|
28 |
-
|
29 |
-
class STFT(torch.nn.Module):
|
30 |
-
def __init__(
|
31 |
-
self, filter_length=1024, hop_length=512, win_length=None, window="hann"
|
32 |
-
):
|
33 |
-
"""
|
34 |
-
This module implements an STFT using 1D convolution and 1D transpose convolutions.
|
35 |
-
This is a bit tricky so there are some cases that probably won't work as working
|
36 |
-
out the same sizes before and after in all overlap add setups is tough. Right now,
|
37 |
-
this code should work with hop lengths that are half the filter length (50% overlap
|
38 |
-
between frames).
|
39 |
-
|
40 |
-
Keyword Arguments:
|
41 |
-
filter_length {int} -- Length of filters used (default: {1024})
|
42 |
-
hop_length {int} -- Hop length of STFT (restrict to 50% overlap between frames) (default: {512})
|
43 |
-
win_length {[type]} -- Length of the window function applied to each frame (if not specified, it
|
44 |
-
equals the filter length). (default: {None})
|
45 |
-
window {str} -- Type of window to use (options are bartlett, hann, hamming, blackman, blackmanharris)
|
46 |
-
(default: {'hann'})
|
47 |
-
"""
|
48 |
-
super(STFT, self).__init__()
|
49 |
-
self.filter_length = filter_length
|
50 |
-
self.hop_length = hop_length
|
51 |
-
self.win_length = win_length if win_length else filter_length
|
52 |
-
self.window = window
|
53 |
-
self.forward_transform = None
|
54 |
-
self.pad_amount = int(self.filter_length / 2)
|
55 |
-
fourier_basis = np.fft.fft(np.eye(self.filter_length))
|
56 |
-
|
57 |
-
cutoff = int((self.filter_length / 2 + 1))
|
58 |
-
fourier_basis = np.vstack(
|
59 |
-
[np.real(fourier_basis[:cutoff, :]), np.imag(fourier_basis[:cutoff, :])]
|
60 |
-
)
|
61 |
-
forward_basis = torch.FloatTensor(fourier_basis)
|
62 |
-
inverse_basis = torch.FloatTensor(np.linalg.pinv(fourier_basis))
|
63 |
-
|
64 |
-
assert filter_length >= self.win_length
|
65 |
-
# get window and zero center pad it to filter_length
|
66 |
-
fft_window = get_window(window, self.win_length, fftbins=True)
|
67 |
-
fft_window = pad_center(fft_window, size=filter_length)
|
68 |
-
fft_window = torch.from_numpy(fft_window).float()
|
69 |
-
|
70 |
-
# window the bases
|
71 |
-
forward_basis *= fft_window
|
72 |
-
inverse_basis = (inverse_basis.T * fft_window).T
|
73 |
-
|
74 |
-
self.register_buffer("forward_basis", forward_basis.float())
|
75 |
-
self.register_buffer("inverse_basis", inverse_basis.float())
|
76 |
-
self.register_buffer("fft_window", fft_window.float())
|
77 |
-
|
78 |
-
def transform(self, input_data, return_phase=False):
|
79 |
-
"""Take input data (audio) to STFT domain.
|
80 |
-
|
81 |
-
Arguments:
|
82 |
-
input_data {tensor} -- Tensor of floats, with shape (num_batch, num_samples)
|
83 |
-
|
84 |
-
Returns:
|
85 |
-
magnitude {tensor} -- Magnitude of STFT with shape (num_batch,
|
86 |
-
num_frequencies, num_frames)
|
87 |
-
phase {tensor} -- Phase of STFT with shape (num_batch,
|
88 |
-
num_frequencies, num_frames)
|
89 |
-
"""
|
90 |
-
input_data = F.pad(
|
91 |
-
input_data,
|
92 |
-
(self.pad_amount, self.pad_amount),
|
93 |
-
mode="reflect",
|
94 |
-
)
|
95 |
-
forward_transform = input_data.unfold(
|
96 |
-
1, self.filter_length, self.hop_length
|
97 |
-
).permute(0, 2, 1)
|
98 |
-
forward_transform = torch.matmul(self.forward_basis, forward_transform)
|
99 |
-
cutoff = int((self.filter_length / 2) + 1)
|
100 |
-
real_part = forward_transform[:, :cutoff, :]
|
101 |
-
imag_part = forward_transform[:, cutoff:, :]
|
102 |
-
magnitude = torch.sqrt(real_part**2 + imag_part**2)
|
103 |
-
if return_phase:
|
104 |
-
phase = torch.atan2(imag_part.data, real_part.data)
|
105 |
-
return magnitude, phase
|
106 |
-
else:
|
107 |
-
return magnitude
|
108 |
-
|
109 |
-
def inverse(self, magnitude, phase):
|
110 |
-
"""Call the inverse STFT (iSTFT), given magnitude and phase tensors produced
|
111 |
-
by the ```transform``` function.
|
112 |
-
|
113 |
-
Arguments:
|
114 |
-
magnitude {tensor} -- Magnitude of STFT with shape (num_batch,
|
115 |
-
num_frequencies, num_frames)
|
116 |
-
phase {tensor} -- Phase of STFT with shape (num_batch,
|
117 |
-
num_frequencies, num_frames)
|
118 |
-
|
119 |
-
Returns:
|
120 |
-
inverse_transform {tensor} -- Reconstructed audio given magnitude and phase. Of
|
121 |
-
shape (num_batch, num_samples)
|
122 |
-
"""
|
123 |
-
cat = torch.cat(
|
124 |
-
[magnitude * torch.cos(phase), magnitude * torch.sin(phase)], dim=1
|
125 |
-
)
|
126 |
-
fold = torch.nn.Fold(
|
127 |
-
output_size=(1, (cat.size(-1) - 1) * self.hop_length + self.filter_length),
|
128 |
-
kernel_size=(1, self.filter_length),
|
129 |
-
stride=(1, self.hop_length),
|
130 |
-
)
|
131 |
-
inverse_transform = torch.matmul(self.inverse_basis, cat)
|
132 |
-
inverse_transform = fold(inverse_transform)[
|
133 |
-
:, 0, 0, self.pad_amount : -self.pad_amount
|
134 |
-
]
|
135 |
-
window_square_sum = (
|
136 |
-
self.fft_window.pow(2).repeat(cat.size(-1), 1).T.unsqueeze(0)
|
137 |
-
)
|
138 |
-
window_square_sum = fold(window_square_sum)[
|
139 |
-
:, 0, 0, self.pad_amount : -self.pad_amount
|
140 |
-
]
|
141 |
-
inverse_transform /= window_square_sum
|
142 |
-
return inverse_transform
|
143 |
-
|
144 |
-
def forward(self, input_data):
|
145 |
-
"""Take input data (audio) to STFT domain and then back to audio.
|
146 |
-
|
147 |
-
Arguments:
|
148 |
-
input_data {tensor} -- Tensor of floats, with shape (num_batch, num_samples)
|
149 |
-
|
150 |
-
Returns:
|
151 |
-
reconstruction {tensor} -- Reconstructed audio given magnitude and phase. Of
|
152 |
-
shape (num_batch, num_samples)
|
153 |
-
"""
|
154 |
-
self.magnitude, self.phase = self.transform(input_data, return_phase=True)
|
155 |
-
reconstruction = self.inverse(self.magnitude, self.phase)
|
156 |
-
return reconstruction
|
157 |
-
|
158 |
-
|
159 |
-
from time import time as ttime
|
160 |
-
|
161 |
-
|
162 |
-
class BiGRU(nn.Module):
|
163 |
-
def __init__(self, input_features, hidden_features, num_layers):
|
164 |
-
super(BiGRU, self).__init__()
|
165 |
-
self.gru = nn.GRU(
|
166 |
-
input_features,
|
167 |
-
hidden_features,
|
168 |
-
num_layers=num_layers,
|
169 |
-
batch_first=True,
|
170 |
-
bidirectional=True,
|
171 |
-
)
|
172 |
-
|
173 |
-
def forward(self, x):
|
174 |
-
return self.gru(x)[0]
|
175 |
-
|
176 |
-
|
177 |
-
class ConvBlockRes(nn.Module):
|
178 |
-
def __init__(self, in_channels, out_channels, momentum=0.01):
|
179 |
-
super(ConvBlockRes, self).__init__()
|
180 |
-
self.conv = nn.Sequential(
|
181 |
-
nn.Conv2d(
|
182 |
-
in_channels=in_channels,
|
183 |
-
out_channels=out_channels,
|
184 |
-
kernel_size=(3, 3),
|
185 |
-
stride=(1, 1),
|
186 |
-
padding=(1, 1),
|
187 |
-
bias=False,
|
188 |
-
),
|
189 |
-
nn.BatchNorm2d(out_channels, momentum=momentum),
|
190 |
-
nn.ReLU(),
|
191 |
-
nn.Conv2d(
|
192 |
-
in_channels=out_channels,
|
193 |
-
out_channels=out_channels,
|
194 |
-
kernel_size=(3, 3),
|
195 |
-
stride=(1, 1),
|
196 |
-
padding=(1, 1),
|
197 |
-
bias=False,
|
198 |
-
),
|
199 |
-
nn.BatchNorm2d(out_channels, momentum=momentum),
|
200 |
-
nn.ReLU(),
|
201 |
-
)
|
202 |
-
# self.shortcut:Optional[nn.Module] = None
|
203 |
-
if in_channels != out_channels:
|
204 |
-
self.shortcut = nn.Conv2d(in_channels, out_channels, (1, 1))
|
205 |
-
|
206 |
-
def forward(self, x: torch.Tensor):
|
207 |
-
if not hasattr(self, "shortcut"):
|
208 |
-
return self.conv(x) + x
|
209 |
-
else:
|
210 |
-
return self.conv(x) + self.shortcut(x)
|
211 |
-
|
212 |
-
|
213 |
-
class Encoder(nn.Module):
|
214 |
-
def __init__(
|
215 |
-
self,
|
216 |
-
in_channels,
|
217 |
-
in_size,
|
218 |
-
n_encoders,
|
219 |
-
kernel_size,
|
220 |
-
n_blocks,
|
221 |
-
out_channels=16,
|
222 |
-
momentum=0.01,
|
223 |
-
):
|
224 |
-
super(Encoder, self).__init__()
|
225 |
-
self.n_encoders = n_encoders
|
226 |
-
self.bn = nn.BatchNorm2d(in_channels, momentum=momentum)
|
227 |
-
self.layers = nn.ModuleList()
|
228 |
-
self.latent_channels = []
|
229 |
-
for i in range(self.n_encoders):
|
230 |
-
self.layers.append(
|
231 |
-
ResEncoderBlock(
|
232 |
-
in_channels, out_channels, kernel_size, n_blocks, momentum=momentum
|
233 |
-
)
|
234 |
-
)
|
235 |
-
self.latent_channels.append([out_channels, in_size])
|
236 |
-
in_channels = out_channels
|
237 |
-
out_channels *= 2
|
238 |
-
in_size //= 2
|
239 |
-
self.out_size = in_size
|
240 |
-
self.out_channel = out_channels
|
241 |
-
|
242 |
-
def forward(self, x: torch.Tensor):
|
243 |
-
concat_tensors: List[torch.Tensor] = []
|
244 |
-
x = self.bn(x)
|
245 |
-
for i, layer in enumerate(self.layers):
|
246 |
-
t, x = layer(x)
|
247 |
-
concat_tensors.append(t)
|
248 |
-
return x, concat_tensors
|
249 |
-
|
250 |
-
|
251 |
-
class ResEncoderBlock(nn.Module):
|
252 |
-
def __init__(
|
253 |
-
self, in_channels, out_channels, kernel_size, n_blocks=1, momentum=0.01
|
254 |
-
):
|
255 |
-
super(ResEncoderBlock, self).__init__()
|
256 |
-
self.n_blocks = n_blocks
|
257 |
-
self.conv = nn.ModuleList()
|
258 |
-
self.conv.append(ConvBlockRes(in_channels, out_channels, momentum))
|
259 |
-
for i in range(n_blocks - 1):
|
260 |
-
self.conv.append(ConvBlockRes(out_channels, out_channels, momentum))
|
261 |
-
self.kernel_size = kernel_size
|
262 |
-
if self.kernel_size is not None:
|
263 |
-
self.pool = nn.AvgPool2d(kernel_size=kernel_size)
|
264 |
-
|
265 |
-
def forward(self, x):
|
266 |
-
for i, conv in enumerate(self.conv):
|
267 |
-
x = conv(x)
|
268 |
-
if self.kernel_size is not None:
|
269 |
-
return x, self.pool(x)
|
270 |
-
else:
|
271 |
-
return x
|
272 |
-
|
273 |
-
|
274 |
-
class Intermediate(nn.Module): #
|
275 |
-
def __init__(self, in_channels, out_channels, n_inters, n_blocks, momentum=0.01):
|
276 |
-
super(Intermediate, self).__init__()
|
277 |
-
self.n_inters = n_inters
|
278 |
-
self.layers = nn.ModuleList()
|
279 |
-
self.layers.append(
|
280 |
-
ResEncoderBlock(in_channels, out_channels, None, n_blocks, momentum)
|
281 |
-
)
|
282 |
-
for i in range(self.n_inters - 1):
|
283 |
-
self.layers.append(
|
284 |
-
ResEncoderBlock(out_channels, out_channels, None, n_blocks, momentum)
|
285 |
-
)
|
286 |
-
|
287 |
-
def forward(self, x):
|
288 |
-
for i, layer in enumerate(self.layers):
|
289 |
-
x = layer(x)
|
290 |
-
return x
|
291 |
-
|
292 |
-
|
293 |
-
class ResDecoderBlock(nn.Module):
|
294 |
-
def __init__(self, in_channels, out_channels, stride, n_blocks=1, momentum=0.01):
|
295 |
-
super(ResDecoderBlock, self).__init__()
|
296 |
-
out_padding = (0, 1) if stride == (1, 2) else (1, 1)
|
297 |
-
self.n_blocks = n_blocks
|
298 |
-
self.conv1 = nn.Sequential(
|
299 |
-
nn.ConvTranspose2d(
|
300 |
-
in_channels=in_channels,
|
301 |
-
out_channels=out_channels,
|
302 |
-
kernel_size=(3, 3),
|
303 |
-
stride=stride,
|
304 |
-
padding=(1, 1),
|
305 |
-
output_padding=out_padding,
|
306 |
-
bias=False,
|
307 |
-
),
|
308 |
-
nn.BatchNorm2d(out_channels, momentum=momentum),
|
309 |
-
nn.ReLU(),
|
310 |
-
)
|
311 |
-
self.conv2 = nn.ModuleList()
|
312 |
-
self.conv2.append(ConvBlockRes(out_channels * 2, out_channels, momentum))
|
313 |
-
for i in range(n_blocks - 1):
|
314 |
-
self.conv2.append(ConvBlockRes(out_channels, out_channels, momentum))
|
315 |
-
|
316 |
-
def forward(self, x, concat_tensor):
|
317 |
-
x = self.conv1(x)
|
318 |
-
x = torch.cat((x, concat_tensor), dim=1)
|
319 |
-
for i, conv2 in enumerate(self.conv2):
|
320 |
-
x = conv2(x)
|
321 |
-
return x
|
322 |
-
|
323 |
-
|
324 |
-
class Decoder(nn.Module):
|
325 |
-
def __init__(self, in_channels, n_decoders, stride, n_blocks, momentum=0.01):
|
326 |
-
super(Decoder, self).__init__()
|
327 |
-
self.layers = nn.ModuleList()
|
328 |
-
self.n_decoders = n_decoders
|
329 |
-
for i in range(self.n_decoders):
|
330 |
-
out_channels = in_channels // 2
|
331 |
-
self.layers.append(
|
332 |
-
ResDecoderBlock(in_channels, out_channels, stride, n_blocks, momentum)
|
333 |
-
)
|
334 |
-
in_channels = out_channels
|
335 |
-
|
336 |
-
def forward(self, x: torch.Tensor, concat_tensors: List[torch.Tensor]):
|
337 |
-
for i, layer in enumerate(self.layers):
|
338 |
-
x = layer(x, concat_tensors[-1 - i])
|
339 |
-
return x
|
340 |
-
|
341 |
-
|
342 |
-
class DeepUnet(nn.Module):
|
343 |
-
def __init__(
|
344 |
-
self,
|
345 |
-
kernel_size,
|
346 |
-
n_blocks,
|
347 |
-
en_de_layers=5,
|
348 |
-
inter_layers=4,
|
349 |
-
in_channels=1,
|
350 |
-
en_out_channels=16,
|
351 |
-
):
|
352 |
-
super(DeepUnet, self).__init__()
|
353 |
-
self.encoder = Encoder(
|
354 |
-
in_channels, 128, en_de_layers, kernel_size, n_blocks, en_out_channels
|
355 |
-
)
|
356 |
-
self.intermediate = Intermediate(
|
357 |
-
self.encoder.out_channel // 2,
|
358 |
-
self.encoder.out_channel,
|
359 |
-
inter_layers,
|
360 |
-
n_blocks,
|
361 |
-
)
|
362 |
-
self.decoder = Decoder(
|
363 |
-
self.encoder.out_channel, en_de_layers, kernel_size, n_blocks
|
364 |
-
)
|
365 |
-
|
366 |
-
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
367 |
-
x, concat_tensors = self.encoder(x)
|
368 |
-
x = self.intermediate(x)
|
369 |
-
x = self.decoder(x, concat_tensors)
|
370 |
-
return x
|
371 |
-
|
372 |
-
|
373 |
-
class E2E(nn.Module):
|
374 |
-
def __init__(
|
375 |
-
self,
|
376 |
-
n_blocks,
|
377 |
-
n_gru,
|
378 |
-
kernel_size,
|
379 |
-
en_de_layers=5,
|
380 |
-
inter_layers=4,
|
381 |
-
in_channels=1,
|
382 |
-
en_out_channels=16,
|
383 |
-
):
|
384 |
-
super(E2E, self).__init__()
|
385 |
-
self.unet = DeepUnet(
|
386 |
-
kernel_size,
|
387 |
-
n_blocks,
|
388 |
-
en_de_layers,
|
389 |
-
inter_layers,
|
390 |
-
in_channels,
|
391 |
-
en_out_channels,
|
392 |
-
)
|
393 |
-
self.cnn = nn.Conv2d(en_out_channels, 3, (3, 3), padding=(1, 1))
|
394 |
-
if n_gru:
|
395 |
-
self.fc = nn.Sequential(
|
396 |
-
BiGRU(3 * 128, 256, n_gru),
|
397 |
-
nn.Linear(512, 360),
|
398 |
-
nn.Dropout(0.25),
|
399 |
-
nn.Sigmoid(),
|
400 |
-
)
|
401 |
-
else:
|
402 |
-
self.fc = nn.Sequential(
|
403 |
-
nn.Linear(3 * nn.N_MELS, nn.N_CLASS), nn.Dropout(0.25), nn.Sigmoid()
|
404 |
-
)
|
405 |
-
|
406 |
-
def forward(self, mel):
|
407 |
-
# print(mel.shape)
|
408 |
-
mel = mel.transpose(-1, -2).unsqueeze(1)
|
409 |
-
x = self.cnn(self.unet(mel)).transpose(1, 2).flatten(-2)
|
410 |
-
x = self.fc(x)
|
411 |
-
# print(x.shape)
|
412 |
-
return x
|
413 |
-
|
414 |
-
|
415 |
-
from librosa.filters import mel
|
416 |
-
|
417 |
-
|
418 |
-
class MelSpectrogram(torch.nn.Module):
|
419 |
-
def __init__(
|
420 |
-
self,
|
421 |
-
is_half,
|
422 |
-
n_mel_channels,
|
423 |
-
sampling_rate,
|
424 |
-
win_length,
|
425 |
-
hop_length,
|
426 |
-
n_fft=None,
|
427 |
-
mel_fmin=0,
|
428 |
-
mel_fmax=None,
|
429 |
-
clamp=1e-5,
|
430 |
-
):
|
431 |
-
super().__init__()
|
432 |
-
n_fft = win_length if n_fft is None else n_fft
|
433 |
-
self.hann_window = {}
|
434 |
-
mel_basis = mel(
|
435 |
-
sr=sampling_rate,
|
436 |
-
n_fft=n_fft,
|
437 |
-
n_mels=n_mel_channels,
|
438 |
-
fmin=mel_fmin,
|
439 |
-
fmax=mel_fmax,
|
440 |
-
htk=True,
|
441 |
-
)
|
442 |
-
mel_basis = torch.from_numpy(mel_basis).float()
|
443 |
-
self.register_buffer("mel_basis", mel_basis)
|
444 |
-
self.n_fft = win_length if n_fft is None else n_fft
|
445 |
-
self.hop_length = hop_length
|
446 |
-
self.win_length = win_length
|
447 |
-
self.sampling_rate = sampling_rate
|
448 |
-
self.n_mel_channels = n_mel_channels
|
449 |
-
self.clamp = clamp
|
450 |
-
self.is_half = is_half
|
451 |
-
|
452 |
-
def forward(self, audio, keyshift=0, speed=1, center=True):
|
453 |
-
factor = 2 ** (keyshift / 12)
|
454 |
-
n_fft_new = int(np.round(self.n_fft * factor))
|
455 |
-
win_length_new = int(np.round(self.win_length * factor))
|
456 |
-
hop_length_new = int(np.round(self.hop_length * speed))
|
457 |
-
keyshift_key = str(keyshift) + "_" + str(audio.device)
|
458 |
-
if keyshift_key not in self.hann_window:
|
459 |
-
self.hann_window[keyshift_key] = torch.hann_window(win_length_new).to(
|
460 |
-
audio.device
|
461 |
-
)
|
462 |
-
if "privateuseone" in str(audio.device):
|
463 |
-
if not hasattr(self, "stft"):
|
464 |
-
self.stft = STFT(
|
465 |
-
filter_length=n_fft_new,
|
466 |
-
hop_length=hop_length_new,
|
467 |
-
win_length=win_length_new,
|
468 |
-
window="hann",
|
469 |
-
).to(audio.device)
|
470 |
-
magnitude = self.stft.transform(audio)
|
471 |
-
else:
|
472 |
-
fft = torch.stft(
|
473 |
-
audio,
|
474 |
-
n_fft=n_fft_new,
|
475 |
-
hop_length=hop_length_new,
|
476 |
-
win_length=win_length_new,
|
477 |
-
window=self.hann_window[keyshift_key],
|
478 |
-
center=center,
|
479 |
-
return_complex=True,
|
480 |
-
)
|
481 |
-
magnitude = torch.sqrt(fft.real.pow(2) + fft.imag.pow(2))
|
482 |
-
if keyshift != 0:
|
483 |
-
size = self.n_fft // 2 + 1
|
484 |
-
resize = magnitude.size(1)
|
485 |
-
if resize < size:
|
486 |
-
magnitude = F.pad(magnitude, (0, 0, 0, size - resize))
|
487 |
-
magnitude = magnitude[:, :size, :] * self.win_length / win_length_new
|
488 |
-
mel_output = torch.matmul(self.mel_basis, magnitude)
|
489 |
-
if self.is_half == True:
|
490 |
-
mel_output = mel_output.half()
|
491 |
-
log_mel_spec = torch.log(torch.clamp(mel_output, min=self.clamp))
|
492 |
-
return log_mel_spec
|
493 |
-
|
494 |
-
|
495 |
-
class RMVPE:
|
496 |
-
def __init__(self, model_path: str, is_half, device=None, use_jit=False):
|
497 |
-
self.resample_kernel = {}
|
498 |
-
self.resample_kernel = {}
|
499 |
-
self.is_half = is_half
|
500 |
-
if device is None:
|
501 |
-
device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
502 |
-
self.device = device
|
503 |
-
self.mel_extractor = MelSpectrogram(
|
504 |
-
is_half, 128, 16000, 1024, 160, None, 30, 8000
|
505 |
-
).to(device)
|
506 |
-
if "privateuseone" in str(device):
|
507 |
-
import onnxruntime as ort
|
508 |
-
|
509 |
-
ort_session = ort.InferenceSession(
|
510 |
-
"%s/rmvpe.onnx" % os.environ["rmvpe_root"],
|
511 |
-
providers=["DmlExecutionProvider"],
|
512 |
-
)
|
513 |
-
self.model = ort_session
|
514 |
-
else:
|
515 |
-
if str(self.device) == "cuda":
|
516 |
-
self.device = torch.device("cuda:0")
|
517 |
-
|
518 |
-
def get_jit_model():
|
519 |
-
jit_model_path = model_path.rstrip(".pth")
|
520 |
-
jit_model_path += ".half.jit" if is_half else ".jit"
|
521 |
-
reload = False
|
522 |
-
if os.path.exists(jit_model_path):
|
523 |
-
ckpt = jit.load(jit_model_path)
|
524 |
-
model_device = ckpt["device"]
|
525 |
-
if model_device != str(self.device):
|
526 |
-
reload = True
|
527 |
-
else:
|
528 |
-
reload = True
|
529 |
-
|
530 |
-
if reload:
|
531 |
-
ckpt = jit.rmvpe_jit_export(
|
532 |
-
model_path=model_path,
|
533 |
-
mode="script",
|
534 |
-
inputs_path=None,
|
535 |
-
save_path=jit_model_path,
|
536 |
-
device=device,
|
537 |
-
is_half=is_half,
|
538 |
-
)
|
539 |
-
model = torch.jit.load(BytesIO(ckpt["model"]), map_location=device)
|
540 |
-
return model
|
541 |
-
|
542 |
-
def get_default_model():
|
543 |
-
model = E2E(4, 1, (2, 2))
|
544 |
-
ckpt = torch.load(model_path, map_location="cpu")
|
545 |
-
model.load_state_dict(ckpt)
|
546 |
-
model.eval()
|
547 |
-
if is_half:
|
548 |
-
model = model.half()
|
549 |
-
else:
|
550 |
-
model = model.float()
|
551 |
-
return model
|
552 |
-
|
553 |
-
if use_jit:
|
554 |
-
if is_half and "cpu" in str(self.device):
|
555 |
-
logger.warning(
|
556 |
-
"Use default rmvpe model. \
|
557 |
-
Jit is not supported on the CPU for half floating point"
|
558 |
-
)
|
559 |
-
self.model = get_default_model()
|
560 |
-
else:
|
561 |
-
self.model = get_jit_model()
|
562 |
-
else:
|
563 |
-
self.model = get_default_model()
|
564 |
-
|
565 |
-
self.model = self.model.to(device)
|
566 |
-
cents_mapping = 20 * np.arange(360) + 1997.3794084376191
|
567 |
-
self.cents_mapping = np.pad(cents_mapping, (4, 4)) # 368
|
568 |
-
|
569 |
-
def mel2hidden(self, mel):
|
570 |
-
with torch.no_grad():
|
571 |
-
n_frames = mel.shape[-1]
|
572 |
-
n_pad = 32 * ((n_frames - 1) // 32 + 1) - n_frames
|
573 |
-
if n_pad > 0:
|
574 |
-
mel = F.pad(mel, (0, n_pad), mode="constant")
|
575 |
-
if "privateuseone" in str(self.device):
|
576 |
-
onnx_input_name = self.model.get_inputs()[0].name
|
577 |
-
onnx_outputs_names = self.model.get_outputs()[0].name
|
578 |
-
hidden = self.model.run(
|
579 |
-
[onnx_outputs_names],
|
580 |
-
input_feed={onnx_input_name: mel.cpu().numpy()},
|
581 |
-
)[0]
|
582 |
-
else:
|
583 |
-
mel = mel.half() if self.is_half else mel.float()
|
584 |
-
hidden = self.model(mel)
|
585 |
-
return hidden[:, :n_frames]
|
586 |
-
|
587 |
-
def decode(self, hidden, thred=0.03):
|
588 |
-
cents_pred = self.to_local_average_cents(hidden, thred=thred)
|
589 |
-
f0 = 10 * (2 ** (cents_pred / 1200))
|
590 |
-
f0[f0 == 10] = 0
|
591 |
-
# f0 = np.array([10 * (2 ** (cent_pred / 1200)) if cent_pred else 0 for cent_pred in cents_pred])
|
592 |
-
return f0
|
593 |
-
|
594 |
-
def infer_from_audio(self, audio, thred=0.03):
|
595 |
-
# torch.cuda.synchronize()
|
596 |
-
# t0 = ttime()
|
597 |
-
if not torch.is_tensor(audio):
|
598 |
-
audio = torch.from_numpy(audio)
|
599 |
-
mel = self.mel_extractor(
|
600 |
-
audio.float().to(self.device).unsqueeze(0), center=True
|
601 |
-
)
|
602 |
-
# print(123123123,mel.device.type)
|
603 |
-
# torch.cuda.synchronize()
|
604 |
-
# t1 = ttime()
|
605 |
-
hidden = self.mel2hidden(mel)
|
606 |
-
# torch.cuda.synchronize()
|
607 |
-
# t2 = ttime()
|
608 |
-
# print(234234,hidden.device.type)
|
609 |
-
if "privateuseone" not in str(self.device):
|
610 |
-
hidden = hidden.squeeze(0).cpu().numpy()
|
611 |
-
else:
|
612 |
-
hidden = hidden[0]
|
613 |
-
if self.is_half == True:
|
614 |
-
hidden = hidden.astype("float32")
|
615 |
-
|
616 |
-
f0 = self.decode(hidden, thred=thred)
|
617 |
-
# torch.cuda.synchronize()
|
618 |
-
# t3 = ttime()
|
619 |
-
# print("hmvpe:%s\t%s\t%s\t%s"%(t1-t0,t2-t1,t3-t2,t3-t0))
|
620 |
-
return f0
|
621 |
-
|
622 |
-
def to_local_average_cents(self, salience, thred=0.05):
|
623 |
-
# t0 = ttime()
|
624 |
-
center = np.argmax(salience, axis=1) # 帧长#index
|
625 |
-
salience = np.pad(salience, ((0, 0), (4, 4))) # 帧长,368
|
626 |
-
# t1 = ttime()
|
627 |
-
center += 4
|
628 |
-
todo_salience = []
|
629 |
-
todo_cents_mapping = []
|
630 |
-
starts = center - 4
|
631 |
-
ends = center + 5
|
632 |
-
for idx in range(salience.shape[0]):
|
633 |
-
todo_salience.append(salience[:, starts[idx] : ends[idx]][idx])
|
634 |
-
todo_cents_mapping.append(self.cents_mapping[starts[idx] : ends[idx]])
|
635 |
-
# t2 = ttime()
|
636 |
-
todo_salience = np.array(todo_salience) # 帧长,9
|
637 |
-
todo_cents_mapping = np.array(todo_cents_mapping) # 帧长,9
|
638 |
-
product_sum = np.sum(todo_salience * todo_cents_mapping, 1)
|
639 |
-
weight_sum = np.sum(todo_salience, 1) # 帧长
|
640 |
-
devided = product_sum / weight_sum # 帧长
|
641 |
-
# t3 = ttime()
|
642 |
-
maxx = np.max(salience, axis=1) # 帧长
|
643 |
-
devided[maxx <= thred] = 0
|
644 |
-
# t4 = ttime()
|
645 |
-
# print("decode:%s\t%s\t%s\t%s" % (t1 - t0, t2 - t1, t3 - t2, t4 - t3))
|
646 |
-
return devided
|
647 |
-
|
648 |
-
|
649 |
-
if __name__ == "__main__":
|
650 |
-
import librosa
|
651 |
-
import soundfile as sf
|
652 |
-
|
653 |
-
audio, sampling_rate = sf.read(r"C:\Users\liujing04\Desktop\Z\冬之花clip1.wav")
|
654 |
-
if len(audio.shape) > 1:
|
655 |
-
audio = librosa.to_mono(audio.transpose(1, 0))
|
656 |
-
audio_bak = audio.copy()
|
657 |
-
if sampling_rate != 16000:
|
658 |
-
audio = librosa.resample(audio, orig_sr=sampling_rate, target_sr=16000)
|
659 |
-
model_path = r"D:\BaiduNetdiskDownload\RVC-beta-v2-0727AMD_realtime\rmvpe.pt"
|
660 |
-
thred = 0.03 # 0.01
|
661 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
662 |
-
rmvpe = RMVPE(model_path, is_half=False, device=device)
|
663 |
-
t0 = ttime()
|
664 |
-
f0 = rmvpe.infer_from_audio(audio, thred=thred)
|
665 |
-
# f0 = rmvpe.infer_from_audio(audio, thred=thred)
|
666 |
-
# f0 = rmvpe.infer_from_audio(audio, thred=thred)
|
667 |
-
# f0 = rmvpe.infer_from_audio(audio, thred=thred)
|
668 |
-
# f0 = rmvpe.infer_from_audio(audio, thred=thred)
|
669 |
-
t1 = ttime()
|
670 |
-
logger.info("%s %.2f", f0.shape, t1 - t0)
|
|
|
1 |
+
from io import BytesIO
|
2 |
+
import os
|
3 |
+
from typing import List, Optional, Tuple
|
4 |
+
import numpy as np
|
5 |
+
import torch
|
6 |
+
|
7 |
+
from libs import jit
|
8 |
+
|
9 |
+
try:
|
10 |
+
# Fix "Torch not compiled with CUDA enabled"
|
11 |
+
import intel_extension_for_pytorch as ipex # pylint: disable=import-error, unused-import
|
12 |
+
|
13 |
+
if torch.xpu.is_available():
|
14 |
+
from infer.modules.ipex import ipex_init
|
15 |
+
|
16 |
+
ipex_init()
|
17 |
+
except Exception: # pylint: disable=broad-exception-caught
|
18 |
+
pass
|
19 |
+
import torch.nn as nn
|
20 |
+
import torch.nn.functional as F
|
21 |
+
from librosa.util import normalize, pad_center, tiny
|
22 |
+
from scipy.signal import get_window
|
23 |
+
|
24 |
+
import logging
|
25 |
+
|
26 |
+
logger = logging.getLogger(__name__)
|
27 |
+
|
28 |
+
|
29 |
+
class STFT(torch.nn.Module):
|
30 |
+
def __init__(
|
31 |
+
self, filter_length=1024, hop_length=512, win_length=None, window="hann"
|
32 |
+
):
|
33 |
+
"""
|
34 |
+
This module implements an STFT using 1D convolution and 1D transpose convolutions.
|
35 |
+
This is a bit tricky so there are some cases that probably won't work as working
|
36 |
+
out the same sizes before and after in all overlap add setups is tough. Right now,
|
37 |
+
this code should work with hop lengths that are half the filter length (50% overlap
|
38 |
+
between frames).
|
39 |
+
|
40 |
+
Keyword Arguments:
|
41 |
+
filter_length {int} -- Length of filters used (default: {1024})
|
42 |
+
hop_length {int} -- Hop length of STFT (restrict to 50% overlap between frames) (default: {512})
|
43 |
+
win_length {[type]} -- Length of the window function applied to each frame (if not specified, it
|
44 |
+
equals the filter length). (default: {None})
|
45 |
+
window {str} -- Type of window to use (options are bartlett, hann, hamming, blackman, blackmanharris)
|
46 |
+
(default: {'hann'})
|
47 |
+
"""
|
48 |
+
super(STFT, self).__init__()
|
49 |
+
self.filter_length = filter_length
|
50 |
+
self.hop_length = hop_length
|
51 |
+
self.win_length = win_length if win_length else filter_length
|
52 |
+
self.window = window
|
53 |
+
self.forward_transform = None
|
54 |
+
self.pad_amount = int(self.filter_length / 2)
|
55 |
+
fourier_basis = np.fft.fft(np.eye(self.filter_length))
|
56 |
+
|
57 |
+
cutoff = int((self.filter_length / 2 + 1))
|
58 |
+
fourier_basis = np.vstack(
|
59 |
+
[np.real(fourier_basis[:cutoff, :]), np.imag(fourier_basis[:cutoff, :])]
|
60 |
+
)
|
61 |
+
forward_basis = torch.FloatTensor(fourier_basis)
|
62 |
+
inverse_basis = torch.FloatTensor(np.linalg.pinv(fourier_basis))
|
63 |
+
|
64 |
+
assert filter_length >= self.win_length
|
65 |
+
# get window and zero center pad it to filter_length
|
66 |
+
fft_window = get_window(window, self.win_length, fftbins=True)
|
67 |
+
fft_window = pad_center(fft_window, size=filter_length)
|
68 |
+
fft_window = torch.from_numpy(fft_window).float()
|
69 |
+
|
70 |
+
# window the bases
|
71 |
+
forward_basis *= fft_window
|
72 |
+
inverse_basis = (inverse_basis.T * fft_window).T
|
73 |
+
|
74 |
+
self.register_buffer("forward_basis", forward_basis.float())
|
75 |
+
self.register_buffer("inverse_basis", inverse_basis.float())
|
76 |
+
self.register_buffer("fft_window", fft_window.float())
|
77 |
+
|
78 |
+
def transform(self, input_data, return_phase=False):
|
79 |
+
"""Take input data (audio) to STFT domain.
|
80 |
+
|
81 |
+
Arguments:
|
82 |
+
input_data {tensor} -- Tensor of floats, with shape (num_batch, num_samples)
|
83 |
+
|
84 |
+
Returns:
|
85 |
+
magnitude {tensor} -- Magnitude of STFT with shape (num_batch,
|
86 |
+
num_frequencies, num_frames)
|
87 |
+
phase {tensor} -- Phase of STFT with shape (num_batch,
|
88 |
+
num_frequencies, num_frames)
|
89 |
+
"""
|
90 |
+
input_data = F.pad(
|
91 |
+
input_data,
|
92 |
+
(self.pad_amount, self.pad_amount),
|
93 |
+
mode="reflect",
|
94 |
+
)
|
95 |
+
forward_transform = input_data.unfold(
|
96 |
+
1, self.filter_length, self.hop_length
|
97 |
+
).permute(0, 2, 1)
|
98 |
+
forward_transform = torch.matmul(self.forward_basis, forward_transform)
|
99 |
+
cutoff = int((self.filter_length / 2) + 1)
|
100 |
+
real_part = forward_transform[:, :cutoff, :]
|
101 |
+
imag_part = forward_transform[:, cutoff:, :]
|
102 |
+
magnitude = torch.sqrt(real_part**2 + imag_part**2)
|
103 |
+
if return_phase:
|
104 |
+
phase = torch.atan2(imag_part.data, real_part.data)
|
105 |
+
return magnitude, phase
|
106 |
+
else:
|
107 |
+
return magnitude
|
108 |
+
|
109 |
+
def inverse(self, magnitude, phase):
|
110 |
+
"""Call the inverse STFT (iSTFT), given magnitude and phase tensors produced
|
111 |
+
by the ```transform``` function.
|
112 |
+
|
113 |
+
Arguments:
|
114 |
+
magnitude {tensor} -- Magnitude of STFT with shape (num_batch,
|
115 |
+
num_frequencies, num_frames)
|
116 |
+
phase {tensor} -- Phase of STFT with shape (num_batch,
|
117 |
+
num_frequencies, num_frames)
|
118 |
+
|
119 |
+
Returns:
|
120 |
+
inverse_transform {tensor} -- Reconstructed audio given magnitude and phase. Of
|
121 |
+
shape (num_batch, num_samples)
|
122 |
+
"""
|
123 |
+
cat = torch.cat(
|
124 |
+
[magnitude * torch.cos(phase), magnitude * torch.sin(phase)], dim=1
|
125 |
+
)
|
126 |
+
fold = torch.nn.Fold(
|
127 |
+
output_size=(1, (cat.size(-1) - 1) * self.hop_length + self.filter_length),
|
128 |
+
kernel_size=(1, self.filter_length),
|
129 |
+
stride=(1, self.hop_length),
|
130 |
+
)
|
131 |
+
inverse_transform = torch.matmul(self.inverse_basis, cat)
|
132 |
+
inverse_transform = fold(inverse_transform)[
|
133 |
+
:, 0, 0, self.pad_amount : -self.pad_amount
|
134 |
+
]
|
135 |
+
window_square_sum = (
|
136 |
+
self.fft_window.pow(2).repeat(cat.size(-1), 1).T.unsqueeze(0)
|
137 |
+
)
|
138 |
+
window_square_sum = fold(window_square_sum)[
|
139 |
+
:, 0, 0, self.pad_amount : -self.pad_amount
|
140 |
+
]
|
141 |
+
inverse_transform /= window_square_sum
|
142 |
+
return inverse_transform
|
143 |
+
|
144 |
+
def forward(self, input_data):
|
145 |
+
"""Take input data (audio) to STFT domain and then back to audio.
|
146 |
+
|
147 |
+
Arguments:
|
148 |
+
input_data {tensor} -- Tensor of floats, with shape (num_batch, num_samples)
|
149 |
+
|
150 |
+
Returns:
|
151 |
+
reconstruction {tensor} -- Reconstructed audio given magnitude and phase. Of
|
152 |
+
shape (num_batch, num_samples)
|
153 |
+
"""
|
154 |
+
self.magnitude, self.phase = self.transform(input_data, return_phase=True)
|
155 |
+
reconstruction = self.inverse(self.magnitude, self.phase)
|
156 |
+
return reconstruction
|
157 |
+
|
158 |
+
|
159 |
+
from time import time as ttime
|
160 |
+
|
161 |
+
|
162 |
+
class BiGRU(nn.Module):
|
163 |
+
def __init__(self, input_features, hidden_features, num_layers):
|
164 |
+
super(BiGRU, self).__init__()
|
165 |
+
self.gru = nn.GRU(
|
166 |
+
input_features,
|
167 |
+
hidden_features,
|
168 |
+
num_layers=num_layers,
|
169 |
+
batch_first=True,
|
170 |
+
bidirectional=True,
|
171 |
+
)
|
172 |
+
|
173 |
+
def forward(self, x):
|
174 |
+
return self.gru(x)[0]
|
175 |
+
|
176 |
+
|
177 |
+
class ConvBlockRes(nn.Module):
|
178 |
+
def __init__(self, in_channels, out_channels, momentum=0.01):
|
179 |
+
super(ConvBlockRes, self).__init__()
|
180 |
+
self.conv = nn.Sequential(
|
181 |
+
nn.Conv2d(
|
182 |
+
in_channels=in_channels,
|
183 |
+
out_channels=out_channels,
|
184 |
+
kernel_size=(3, 3),
|
185 |
+
stride=(1, 1),
|
186 |
+
padding=(1, 1),
|
187 |
+
bias=False,
|
188 |
+
),
|
189 |
+
nn.BatchNorm2d(out_channels, momentum=momentum),
|
190 |
+
nn.ReLU(),
|
191 |
+
nn.Conv2d(
|
192 |
+
in_channels=out_channels,
|
193 |
+
out_channels=out_channels,
|
194 |
+
kernel_size=(3, 3),
|
195 |
+
stride=(1, 1),
|
196 |
+
padding=(1, 1),
|
197 |
+
bias=False,
|
198 |
+
),
|
199 |
+
nn.BatchNorm2d(out_channels, momentum=momentum),
|
200 |
+
nn.ReLU(),
|
201 |
+
)
|
202 |
+
# self.shortcut:Optional[nn.Module] = None
|
203 |
+
if in_channels != out_channels:
|
204 |
+
self.shortcut = nn.Conv2d(in_channels, out_channels, (1, 1))
|
205 |
+
|
206 |
+
def forward(self, x: torch.Tensor):
|
207 |
+
if not hasattr(self, "shortcut"):
|
208 |
+
return self.conv(x) + x
|
209 |
+
else:
|
210 |
+
return self.conv(x) + self.shortcut(x)
|
211 |
+
|
212 |
+
|
213 |
+
class Encoder(nn.Module):
|
214 |
+
def __init__(
|
215 |
+
self,
|
216 |
+
in_channels,
|
217 |
+
in_size,
|
218 |
+
n_encoders,
|
219 |
+
kernel_size,
|
220 |
+
n_blocks,
|
221 |
+
out_channels=16,
|
222 |
+
momentum=0.01,
|
223 |
+
):
|
224 |
+
super(Encoder, self).__init__()
|
225 |
+
self.n_encoders = n_encoders
|
226 |
+
self.bn = nn.BatchNorm2d(in_channels, momentum=momentum)
|
227 |
+
self.layers = nn.ModuleList()
|
228 |
+
self.latent_channels = []
|
229 |
+
for i in range(self.n_encoders):
|
230 |
+
self.layers.append(
|
231 |
+
ResEncoderBlock(
|
232 |
+
in_channels, out_channels, kernel_size, n_blocks, momentum=momentum
|
233 |
+
)
|
234 |
+
)
|
235 |
+
self.latent_channels.append([out_channels, in_size])
|
236 |
+
in_channels = out_channels
|
237 |
+
out_channels *= 2
|
238 |
+
in_size //= 2
|
239 |
+
self.out_size = in_size
|
240 |
+
self.out_channel = out_channels
|
241 |
+
|
242 |
+
def forward(self, x: torch.Tensor):
|
243 |
+
concat_tensors: List[torch.Tensor] = []
|
244 |
+
x = self.bn(x)
|
245 |
+
for i, layer in enumerate(self.layers):
|
246 |
+
t, x = layer(x)
|
247 |
+
concat_tensors.append(t)
|
248 |
+
return x, concat_tensors
|
249 |
+
|
250 |
+
|
251 |
+
class ResEncoderBlock(nn.Module):
|
252 |
+
def __init__(
|
253 |
+
self, in_channels, out_channels, kernel_size, n_blocks=1, momentum=0.01
|
254 |
+
):
|
255 |
+
super(ResEncoderBlock, self).__init__()
|
256 |
+
self.n_blocks = n_blocks
|
257 |
+
self.conv = nn.ModuleList()
|
258 |
+
self.conv.append(ConvBlockRes(in_channels, out_channels, momentum))
|
259 |
+
for i in range(n_blocks - 1):
|
260 |
+
self.conv.append(ConvBlockRes(out_channels, out_channels, momentum))
|
261 |
+
self.kernel_size = kernel_size
|
262 |
+
if self.kernel_size is not None:
|
263 |
+
self.pool = nn.AvgPool2d(kernel_size=kernel_size)
|
264 |
+
|
265 |
+
def forward(self, x):
|
266 |
+
for i, conv in enumerate(self.conv):
|
267 |
+
x = conv(x)
|
268 |
+
if self.kernel_size is not None:
|
269 |
+
return x, self.pool(x)
|
270 |
+
else:
|
271 |
+
return x
|
272 |
+
|
273 |
+
|
274 |
+
class Intermediate(nn.Module): #
|
275 |
+
def __init__(self, in_channels, out_channels, n_inters, n_blocks, momentum=0.01):
|
276 |
+
super(Intermediate, self).__init__()
|
277 |
+
self.n_inters = n_inters
|
278 |
+
self.layers = nn.ModuleList()
|
279 |
+
self.layers.append(
|
280 |
+
ResEncoderBlock(in_channels, out_channels, None, n_blocks, momentum)
|
281 |
+
)
|
282 |
+
for i in range(self.n_inters - 1):
|
283 |
+
self.layers.append(
|
284 |
+
ResEncoderBlock(out_channels, out_channels, None, n_blocks, momentum)
|
285 |
+
)
|
286 |
+
|
287 |
+
def forward(self, x):
|
288 |
+
for i, layer in enumerate(self.layers):
|
289 |
+
x = layer(x)
|
290 |
+
return x
|
291 |
+
|
292 |
+
|
293 |
+
class ResDecoderBlock(nn.Module):
|
294 |
+
def __init__(self, in_channels, out_channels, stride, n_blocks=1, momentum=0.01):
|
295 |
+
super(ResDecoderBlock, self).__init__()
|
296 |
+
out_padding = (0, 1) if stride == (1, 2) else (1, 1)
|
297 |
+
self.n_blocks = n_blocks
|
298 |
+
self.conv1 = nn.Sequential(
|
299 |
+
nn.ConvTranspose2d(
|
300 |
+
in_channels=in_channels,
|
301 |
+
out_channels=out_channels,
|
302 |
+
kernel_size=(3, 3),
|
303 |
+
stride=stride,
|
304 |
+
padding=(1, 1),
|
305 |
+
output_padding=out_padding,
|
306 |
+
bias=False,
|
307 |
+
),
|
308 |
+
nn.BatchNorm2d(out_channels, momentum=momentum),
|
309 |
+
nn.ReLU(),
|
310 |
+
)
|
311 |
+
self.conv2 = nn.ModuleList()
|
312 |
+
self.conv2.append(ConvBlockRes(out_channels * 2, out_channels, momentum))
|
313 |
+
for i in range(n_blocks - 1):
|
314 |
+
self.conv2.append(ConvBlockRes(out_channels, out_channels, momentum))
|
315 |
+
|
316 |
+
def forward(self, x, concat_tensor):
|
317 |
+
x = self.conv1(x)
|
318 |
+
x = torch.cat((x, concat_tensor), dim=1)
|
319 |
+
for i, conv2 in enumerate(self.conv2):
|
320 |
+
x = conv2(x)
|
321 |
+
return x
|
322 |
+
|
323 |
+
|
324 |
+
class Decoder(nn.Module):
|
325 |
+
def __init__(self, in_channels, n_decoders, stride, n_blocks, momentum=0.01):
|
326 |
+
super(Decoder, self).__init__()
|
327 |
+
self.layers = nn.ModuleList()
|
328 |
+
self.n_decoders = n_decoders
|
329 |
+
for i in range(self.n_decoders):
|
330 |
+
out_channels = in_channels // 2
|
331 |
+
self.layers.append(
|
332 |
+
ResDecoderBlock(in_channels, out_channels, stride, n_blocks, momentum)
|
333 |
+
)
|
334 |
+
in_channels = out_channels
|
335 |
+
|
336 |
+
def forward(self, x: torch.Tensor, concat_tensors: List[torch.Tensor]):
|
337 |
+
for i, layer in enumerate(self.layers):
|
338 |
+
x = layer(x, concat_tensors[-1 - i])
|
339 |
+
return x
|
340 |
+
|
341 |
+
|
342 |
+
class DeepUnet(nn.Module):
|
343 |
+
def __init__(
|
344 |
+
self,
|
345 |
+
kernel_size,
|
346 |
+
n_blocks,
|
347 |
+
en_de_layers=5,
|
348 |
+
inter_layers=4,
|
349 |
+
in_channels=1,
|
350 |
+
en_out_channels=16,
|
351 |
+
):
|
352 |
+
super(DeepUnet, self).__init__()
|
353 |
+
self.encoder = Encoder(
|
354 |
+
in_channels, 128, en_de_layers, kernel_size, n_blocks, en_out_channels
|
355 |
+
)
|
356 |
+
self.intermediate = Intermediate(
|
357 |
+
self.encoder.out_channel // 2,
|
358 |
+
self.encoder.out_channel,
|
359 |
+
inter_layers,
|
360 |
+
n_blocks,
|
361 |
+
)
|
362 |
+
self.decoder = Decoder(
|
363 |
+
self.encoder.out_channel, en_de_layers, kernel_size, n_blocks
|
364 |
+
)
|
365 |
+
|
366 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
367 |
+
x, concat_tensors = self.encoder(x)
|
368 |
+
x = self.intermediate(x)
|
369 |
+
x = self.decoder(x, concat_tensors)
|
370 |
+
return x
|
371 |
+
|
372 |
+
|
373 |
+
class E2E(nn.Module):
|
374 |
+
def __init__(
|
375 |
+
self,
|
376 |
+
n_blocks,
|
377 |
+
n_gru,
|
378 |
+
kernel_size,
|
379 |
+
en_de_layers=5,
|
380 |
+
inter_layers=4,
|
381 |
+
in_channels=1,
|
382 |
+
en_out_channels=16,
|
383 |
+
):
|
384 |
+
super(E2E, self).__init__()
|
385 |
+
self.unet = DeepUnet(
|
386 |
+
kernel_size,
|
387 |
+
n_blocks,
|
388 |
+
en_de_layers,
|
389 |
+
inter_layers,
|
390 |
+
in_channels,
|
391 |
+
en_out_channels,
|
392 |
+
)
|
393 |
+
self.cnn = nn.Conv2d(en_out_channels, 3, (3, 3), padding=(1, 1))
|
394 |
+
if n_gru:
|
395 |
+
self.fc = nn.Sequential(
|
396 |
+
BiGRU(3 * 128, 256, n_gru),
|
397 |
+
nn.Linear(512, 360),
|
398 |
+
nn.Dropout(0.25),
|
399 |
+
nn.Sigmoid(),
|
400 |
+
)
|
401 |
+
else:
|
402 |
+
self.fc = nn.Sequential(
|
403 |
+
nn.Linear(3 * nn.N_MELS, nn.N_CLASS), nn.Dropout(0.25), nn.Sigmoid()
|
404 |
+
)
|
405 |
+
|
406 |
+
def forward(self, mel):
|
407 |
+
# print(mel.shape)
|
408 |
+
mel = mel.transpose(-1, -2).unsqueeze(1)
|
409 |
+
x = self.cnn(self.unet(mel)).transpose(1, 2).flatten(-2)
|
410 |
+
x = self.fc(x)
|
411 |
+
# print(x.shape)
|
412 |
+
return x
|
413 |
+
|
414 |
+
|
415 |
+
from librosa.filters import mel
|
416 |
+
|
417 |
+
|
418 |
+
class MelSpectrogram(torch.nn.Module):
|
419 |
+
def __init__(
|
420 |
+
self,
|
421 |
+
is_half,
|
422 |
+
n_mel_channels,
|
423 |
+
sampling_rate,
|
424 |
+
win_length,
|
425 |
+
hop_length,
|
426 |
+
n_fft=None,
|
427 |
+
mel_fmin=0,
|
428 |
+
mel_fmax=None,
|
429 |
+
clamp=1e-5,
|
430 |
+
):
|
431 |
+
super().__init__()
|
432 |
+
n_fft = win_length if n_fft is None else n_fft
|
433 |
+
self.hann_window = {}
|
434 |
+
mel_basis = mel(
|
435 |
+
sr=sampling_rate,
|
436 |
+
n_fft=n_fft,
|
437 |
+
n_mels=n_mel_channels,
|
438 |
+
fmin=mel_fmin,
|
439 |
+
fmax=mel_fmax,
|
440 |
+
htk=True,
|
441 |
+
)
|
442 |
+
mel_basis = torch.from_numpy(mel_basis).float()
|
443 |
+
self.register_buffer("mel_basis", mel_basis)
|
444 |
+
self.n_fft = win_length if n_fft is None else n_fft
|
445 |
+
self.hop_length = hop_length
|
446 |
+
self.win_length = win_length
|
447 |
+
self.sampling_rate = sampling_rate
|
448 |
+
self.n_mel_channels = n_mel_channels
|
449 |
+
self.clamp = clamp
|
450 |
+
self.is_half = is_half
|
451 |
+
|
452 |
+
def forward(self, audio, keyshift=0, speed=1, center=True):
|
453 |
+
factor = 2 ** (keyshift / 12)
|
454 |
+
n_fft_new = int(np.round(self.n_fft * factor))
|
455 |
+
win_length_new = int(np.round(self.win_length * factor))
|
456 |
+
hop_length_new = int(np.round(self.hop_length * speed))
|
457 |
+
keyshift_key = str(keyshift) + "_" + str(audio.device)
|
458 |
+
if keyshift_key not in self.hann_window:
|
459 |
+
self.hann_window[keyshift_key] = torch.hann_window(win_length_new).to(
|
460 |
+
audio.device
|
461 |
+
)
|
462 |
+
if "privateuseone" in str(audio.device):
|
463 |
+
if not hasattr(self, "stft"):
|
464 |
+
self.stft = STFT(
|
465 |
+
filter_length=n_fft_new,
|
466 |
+
hop_length=hop_length_new,
|
467 |
+
win_length=win_length_new,
|
468 |
+
window="hann",
|
469 |
+
).to(audio.device)
|
470 |
+
magnitude = self.stft.transform(audio)
|
471 |
+
else:
|
472 |
+
fft = torch.stft(
|
473 |
+
audio,
|
474 |
+
n_fft=n_fft_new,
|
475 |
+
hop_length=hop_length_new,
|
476 |
+
win_length=win_length_new,
|
477 |
+
window=self.hann_window[keyshift_key],
|
478 |
+
center=center,
|
479 |
+
return_complex=True,
|
480 |
+
)
|
481 |
+
magnitude = torch.sqrt(fft.real.pow(2) + fft.imag.pow(2))
|
482 |
+
if keyshift != 0:
|
483 |
+
size = self.n_fft // 2 + 1
|
484 |
+
resize = magnitude.size(1)
|
485 |
+
if resize < size:
|
486 |
+
magnitude = F.pad(magnitude, (0, 0, 0, size - resize))
|
487 |
+
magnitude = magnitude[:, :size, :] * self.win_length / win_length_new
|
488 |
+
mel_output = torch.matmul(self.mel_basis, magnitude)
|
489 |
+
if self.is_half == True:
|
490 |
+
mel_output = mel_output.half()
|
491 |
+
log_mel_spec = torch.log(torch.clamp(mel_output, min=self.clamp))
|
492 |
+
return log_mel_spec
|
493 |
+
|
494 |
+
|
495 |
+
class RMVPE:
|
496 |
+
def __init__(self, model_path: str, is_half, device=None, use_jit=False):
|
497 |
+
self.resample_kernel = {}
|
498 |
+
self.resample_kernel = {}
|
499 |
+
self.is_half = is_half
|
500 |
+
if device is None:
|
501 |
+
device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
502 |
+
self.device = device
|
503 |
+
self.mel_extractor = MelSpectrogram(
|
504 |
+
is_half, 128, 16000, 1024, 160, None, 30, 8000
|
505 |
+
).to(device)
|
506 |
+
if "privateuseone" in str(device):
|
507 |
+
import onnxruntime as ort
|
508 |
+
|
509 |
+
ort_session = ort.InferenceSession(
|
510 |
+
"%s/rmvpe.onnx" % os.environ["rmvpe_root"],
|
511 |
+
providers=["DmlExecutionProvider"],
|
512 |
+
)
|
513 |
+
self.model = ort_session
|
514 |
+
else:
|
515 |
+
if str(self.device) == "cuda":
|
516 |
+
self.device = torch.device("cuda:0")
|
517 |
+
|
518 |
+
def get_jit_model():
|
519 |
+
jit_model_path = model_path.rstrip(".pth")
|
520 |
+
jit_model_path += ".half.jit" if is_half else ".jit"
|
521 |
+
reload = False
|
522 |
+
if os.path.exists(jit_model_path):
|
523 |
+
ckpt = jit.load(jit_model_path)
|
524 |
+
model_device = ckpt["device"]
|
525 |
+
if model_device != str(self.device):
|
526 |
+
reload = True
|
527 |
+
else:
|
528 |
+
reload = True
|
529 |
+
|
530 |
+
if reload:
|
531 |
+
ckpt = jit.rmvpe_jit_export(
|
532 |
+
model_path=model_path,
|
533 |
+
mode="script",
|
534 |
+
inputs_path=None,
|
535 |
+
save_path=jit_model_path,
|
536 |
+
device=device,
|
537 |
+
is_half=is_half,
|
538 |
+
)
|
539 |
+
model = torch.jit.load(BytesIO(ckpt["model"]), map_location=device)
|
540 |
+
return model
|
541 |
+
|
542 |
+
def get_default_model():
|
543 |
+
model = E2E(4, 1, (2, 2))
|
544 |
+
ckpt = torch.load(model_path, map_location="cpu")
|
545 |
+
model.load_state_dict(ckpt)
|
546 |
+
model.eval()
|
547 |
+
if is_half:
|
548 |
+
model = model.half()
|
549 |
+
else:
|
550 |
+
model = model.float()
|
551 |
+
return model
|
552 |
+
|
553 |
+
if use_jit:
|
554 |
+
if is_half and "cpu" in str(self.device):
|
555 |
+
logger.warning(
|
556 |
+
"Use default rmvpe model. \
|
557 |
+
Jit is not supported on the CPU for half floating point"
|
558 |
+
)
|
559 |
+
self.model = get_default_model()
|
560 |
+
else:
|
561 |
+
self.model = get_jit_model()
|
562 |
+
else:
|
563 |
+
self.model = get_default_model()
|
564 |
+
|
565 |
+
self.model = self.model.to(device)
|
566 |
+
cents_mapping = 20 * np.arange(360) + 1997.3794084376191
|
567 |
+
self.cents_mapping = np.pad(cents_mapping, (4, 4)) # 368
|
568 |
+
|
569 |
+
def mel2hidden(self, mel):
|
570 |
+
with torch.no_grad():
|
571 |
+
n_frames = mel.shape[-1]
|
572 |
+
n_pad = 32 * ((n_frames - 1) // 32 + 1) - n_frames
|
573 |
+
if n_pad > 0:
|
574 |
+
mel = F.pad(mel, (0, n_pad), mode="constant")
|
575 |
+
if "privateuseone" in str(self.device):
|
576 |
+
onnx_input_name = self.model.get_inputs()[0].name
|
577 |
+
onnx_outputs_names = self.model.get_outputs()[0].name
|
578 |
+
hidden = self.model.run(
|
579 |
+
[onnx_outputs_names],
|
580 |
+
input_feed={onnx_input_name: mel.cpu().numpy()},
|
581 |
+
)[0]
|
582 |
+
else:
|
583 |
+
mel = mel.half() if self.is_half else mel.float()
|
584 |
+
hidden = self.model(mel)
|
585 |
+
return hidden[:, :n_frames]
|
586 |
+
|
587 |
+
def decode(self, hidden, thred=0.03):
|
588 |
+
cents_pred = self.to_local_average_cents(hidden, thred=thred)
|
589 |
+
f0 = 10 * (2 ** (cents_pred / 1200))
|
590 |
+
f0[f0 == 10] = 0
|
591 |
+
# f0 = np.array([10 * (2 ** (cent_pred / 1200)) if cent_pred else 0 for cent_pred in cents_pred])
|
592 |
+
return f0
|
593 |
+
|
594 |
+
def infer_from_audio(self, audio, thred=0.03):
|
595 |
+
# torch.cuda.synchronize()
|
596 |
+
# t0 = ttime()
|
597 |
+
if not torch.is_tensor(audio):
|
598 |
+
audio = torch.from_numpy(audio)
|
599 |
+
mel = self.mel_extractor(
|
600 |
+
audio.float().to(self.device).unsqueeze(0), center=True
|
601 |
+
)
|
602 |
+
# print(123123123,mel.device.type)
|
603 |
+
# torch.cuda.synchronize()
|
604 |
+
# t1 = ttime()
|
605 |
+
hidden = self.mel2hidden(mel)
|
606 |
+
# torch.cuda.synchronize()
|
607 |
+
# t2 = ttime()
|
608 |
+
# print(234234,hidden.device.type)
|
609 |
+
if "privateuseone" not in str(self.device):
|
610 |
+
hidden = hidden.squeeze(0).cpu().numpy()
|
611 |
+
else:
|
612 |
+
hidden = hidden[0]
|
613 |
+
if self.is_half == True:
|
614 |
+
hidden = hidden.astype("float32")
|
615 |
+
|
616 |
+
f0 = self.decode(hidden, thred=thred)
|
617 |
+
# torch.cuda.synchronize()
|
618 |
+
# t3 = ttime()
|
619 |
+
# print("hmvpe:%s\t%s\t%s\t%s"%(t1-t0,t2-t1,t3-t2,t3-t0))
|
620 |
+
return f0
|
621 |
+
|
622 |
+
def to_local_average_cents(self, salience, thred=0.05):
|
623 |
+
# t0 = ttime()
|
624 |
+
center = np.argmax(salience, axis=1) # 帧长#index
|
625 |
+
salience = np.pad(salience, ((0, 0), (4, 4))) # 帧长,368
|
626 |
+
# t1 = ttime()
|
627 |
+
center += 4
|
628 |
+
todo_salience = []
|
629 |
+
todo_cents_mapping = []
|
630 |
+
starts = center - 4
|
631 |
+
ends = center + 5
|
632 |
+
for idx in range(salience.shape[0]):
|
633 |
+
todo_salience.append(salience[:, starts[idx] : ends[idx]][idx])
|
634 |
+
todo_cents_mapping.append(self.cents_mapping[starts[idx] : ends[idx]])
|
635 |
+
# t2 = ttime()
|
636 |
+
todo_salience = np.array(todo_salience) # 帧长,9
|
637 |
+
todo_cents_mapping = np.array(todo_cents_mapping) # 帧长,9
|
638 |
+
product_sum = np.sum(todo_salience * todo_cents_mapping, 1)
|
639 |
+
weight_sum = np.sum(todo_salience, 1) # 帧长
|
640 |
+
devided = product_sum / weight_sum # 帧长
|
641 |
+
# t3 = ttime()
|
642 |
+
maxx = np.max(salience, axis=1) # 帧长
|
643 |
+
devided[maxx <= thred] = 0
|
644 |
+
# t4 = ttime()
|
645 |
+
# print("decode:%s\t%s\t%s\t%s" % (t1 - t0, t2 - t1, t3 - t2, t4 - t3))
|
646 |
+
return devided
|
647 |
+
|
648 |
+
|
649 |
+
if __name__ == "__main__":
|
650 |
+
import librosa
|
651 |
+
import soundfile as sf
|
652 |
+
|
653 |
+
audio, sampling_rate = sf.read(r"C:\Users\liujing04\Desktop\Z\冬之花clip1.wav")
|
654 |
+
if len(audio.shape) > 1:
|
655 |
+
audio = librosa.to_mono(audio.transpose(1, 0))
|
656 |
+
audio_bak = audio.copy()
|
657 |
+
if sampling_rate != 16000:
|
658 |
+
audio = librosa.resample(audio, orig_sr=sampling_rate, target_sr=16000)
|
659 |
+
model_path = r"D:\BaiduNetdiskDownload\RVC-beta-v2-0727AMD_realtime\rmvpe.pt"
|
660 |
+
thred = 0.03 # 0.01
|
661 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
662 |
+
rmvpe = RMVPE(model_path, is_half=False, device=device)
|
663 |
+
t0 = ttime()
|
664 |
+
f0 = rmvpe.infer_from_audio(audio, thred=thred)
|
665 |
+
# f0 = rmvpe.infer_from_audio(audio, thred=thred)
|
666 |
+
# f0 = rmvpe.infer_from_audio(audio, thred=thred)
|
667 |
+
# f0 = rmvpe.infer_from_audio(audio, thred=thred)
|
668 |
+
# f0 = rmvpe.infer_from_audio(audio, thred=thred)
|
669 |
+
t1 = ttime()
|
670 |
+
logger.info("%s %.2f", f0.shape, t1 - t0)
|